Every instance’s CPU and volume metrics go to its project in Observability,
free and automatically: there’s nothing to install or turn on, and they aren’t billed or
counted against Observability’s limits. They arrive every 60 seconds, in the project’s
production environment, with job="ezgh.compute" and the instance’s ID as instance.
Metrics
| Metric (PromQL) | Labels | Meaning |
|---|---|---|
compute_instance_cpu_utilization_ratio |
Share of the instance’s vCPU capacity used over the minute, 0 to 1 | |
compute_instance_cpu_time_seconds_total |
Time the instance’s vCPUs ran | |
compute_instance_cpu_wait_seconds_total |
Time its vCPUs were ready to run but waited for a host CPU (steal) | |
compute_instance_storage_throttled_seconds_total |
Time its volumes together were at the instance type’s storage limit | |
compute_instance_status_check_failed |
check (system, instance) |
1 when the status check failed, 0 when it passed; nothing while it’s initializing or not applicable |
blockstore_volume_operations_total |
ezgh_volume_id, direction (read, write) |
Operations the volume completed |
blockstore_volume_bytes_total |
ezgh_volume_id, direction |
Bytes it read or wrote |
blockstore_volume_latency_seconds |
ezgh_volume_id, direction |
Average completion time over the minute, when there were operations |
blockstore_volume_throttled_seconds_total |
ezgh_volume_id |
Time the volume was at its IOPS or throughput limit |
The OTLP names, as sent, are compute.instance.cpu.utilization, compute.instance.cpu.time,
compute.instance.cpu.wait, compute.instance.storage.throttled, compute.instance.status_check.failed, blockstore.volume.operations,
blockstore.volume.bytes, blockstore.volume.latency and blockstore.volume.throttled; see
Names and labels.
The metrics cover running instances, and the volumes attached to them. Memory and disk usage inside the guest aren’t measured from outside it: send those yourself, for example with the OpenTelemetry Collector’s host metrics receiver.
Queries
# CPU utilization of one instance
compute_instance_cpu_utilization_ratio{instance="ins_7k2m9x0a4qpe"}
# Instances above 90% CPU
compute_instance_cpu_utilization_ratio > 0.9
# A volume's IOPS, by direction
sum by (direction) (rate(blockstore_volume_operations_total{ezgh_volume_id="vol_4k2m9x0a7qpe"}[5m]))
# A volume's read and write throughput, in bytes per second
sum by (direction) (rate(blockstore_volume_bytes_total{ezgh_volume_id="vol_4k2m9x0a7qpe"}[5m]))
# Volumes that spent time at their limit in the last hour
increase(blockstore_volume_throttled_seconds_total[1h]) > 0
# Instances whose instance status check failed (alert on this for 5 minutes)
max by (instance) (compute_instance_status_check_failed{check="instance"}) == 1Time spent at a limit means the volume or instance needed more than it’s provisioned with:
modify the volume for more IOPS or throughput, or
resize the instance for a higher storage limit. Steal
(cpu_wait) on a shared instance type means its vCPUs waited for host CPUs.
Query them in Observability, build dashboards and set alerts on them like any other metrics.